juejin.cn
60 articles · May 4, 2026 – May 7, 2026
OpenClaw Joins Feishu: AI Agents Shift from Geek Toys to Enterprise Coworkers
OpenClaw on Feishu: AI Agent bottlenecks shift to "where it lives"—workflow embedding beats standalone apps, but data compliance and platform rivalry
Korean Temple Ordains Robot Monk — AI Spectacle Is the Real Bubble Risk
A 130cm robot "ordained" at a Korean temple exposes regressive AI deployment logic. Soulless spectacles drain public trust and fuel the real AI narrat
Stop Scoring RAG by Feel: AI Apps Enter Data-Driven Operations Era
RAGAS uses 4 quantitative metrics to score RAG systems, solving the "feels right but can't prove it" pain point. This marks enterprise AI shifting fro
OpenAI Enforces Phone Verification as Bulk Codex Farming Triggers Risk Control
OpenAI forces SMS verification on ChatGPT/Codex as bots farm free quotas. SMS platforms collapse, normal users suffer. Anti-cheat upgrade, not complia
Self-Attention Powers AI Context — But Few Firms Truly Understand It
Self-attention is the core of mainstream AI, enabling simultaneous word relationship analysis. Understanding it is key to evaluating AI costs and ROI.
Xiaomi MiMo Wastes 6x Compute on Junk Code; LLMs Shift to Delivery Efficiency
Xiaomi MiMo burned 6x compute for junk code while DeepSeek excelled. Benchmarks no longer reflect true dev capability; focus on delivery and costs.
WPS Multidimensional Table Runs Python: Kingsoft Quietly Pivots to Platform
WPS Multidimensional Table adds Python, MCP, and 70+ APIs. Kingsoft pivots to a developer platform, but near-zero AI buzz leaves its ecosystem prospec
OpenClaw Hits 367K Stars: Personal AI Gateways Are Taking Over Your Chat Apps
OpenClaw, a local cross-chat AI gateway, hit 367K GitHub stars. AI entry points are shifting from dedicated webpages to existing chat boxes—a logic sh
OpenClaw Debuts Telegram: AI Agents Escape Chatboxes, Embed in Your IM
OpenClaw connects Telegram first; 30+ IM platforms like Feishu, WeCom ahead. AI moves from chatboxes to daily workflows as on-call digital workers.
LangChain: AI Agents Load Skills On-Demand — Modular Dev Is the New Agent Paradigm
LangChain DeepAgent: AI agents load skill modules on-demand like humans, shifting Agent development from monolithic to pluggable composition for custo
Doubao Agent Introduces Background Tasks: AI Needs Parallel Processing to Ship
Doubao Agent tackles single-thread blocking by adding background tasks. The AI Agent bottleneck is shifting from model capability to engineering archi
Transformer Book Read 3 Times: LLM Race Shifts from API Calls to Foundational Logic
A deep learning book read 3 times. While most only call LLM APIs, understanding principles like attention mechanisms now dictates AI app success and c
DeepSeek-TUI Tops GitHub at 2434 Stars: Terminal AI Agent Goes Practical
Terminal AI agent DeepSeek-TUI installs in 15s, supports MCP & sub-agents. Terminal AI crosses from "works" to "works well"; Chinese LLMs now compete
C++20 Double Buffering Ends Data Queuing: Underlying Engineering Sets AI Limits
C++20 lock-free double buffering doubles memory to parallelize data generation and processing. As LLMs surge, it eliminates idle cycles caused by data
AI Coding Assistants Embed IDEs as Full-Stack Toolchain Competition Intensifies
HagiCode builds code-server across 3 OSs with OmniRoute. AI assistants evolve from chat windows to full IDEs, signaling a push for AI vendor flexibili
Tech Workers Build AI to Socialize for Them: Classic Side-Project Dilemma
Engineers built ClawReach on OpenClaw: AI chats first, humans meet later. Tech done, ops zero. The classic side-project dilemma: can build, can't prom
Stop Guessing RAG Quality: RAGAS Uses AI to Grade AI
RAG quality often relies on guesswork. RAGAS uses 4 metrics and LLM-as-Judge to turn gut feelings into engineering KPIs—vital for enterprise knowledge
OpenAI Codex /goal Command: Unattended Long-Task AI Coding Arrives
OpenAI adds /goal to Codex CLI for unattended continuous task execution. AI coding shifts from Q&A to goal-driven work, but cost overruns and drift ri
LangChain DeepAgents v2 Streams Progress — Opaque Agents Have No Commercial Value
LangChain updates DeepAgents streaming, solving multi-agent black-screen waits. We judge: real-time AI transparency is make-or-break for user retentio
LangChain's Context Engineering: Cramming AI With Data Makes It Dumber
More data makes LLMs dumber. LangChain's Context Engineering systematically manages AI's "field of view," marking a shift from parameter rivalry to en
OpenClaw Integrates Feishu: AI Agents Finally Join the Corporate Address Book
OpenClaw integrates Feishu, shifting open-source Agents from geek toys to group members handling daily collaboration in mainstream workflows—a pivot i
Palantir Wins Enterprise AI With 20-Year-Old Design: Data Structure Beats Models
Palantir wins via 20-year-old Ontology, not models. Enterprise AI's last-mile block is data lacking business semantics, shifting the competitive focus
AI Rewrites Open Source With Just a Dependency List — Licenses Officially Dead
Malus.sh rewrites open source into legally distinct code, bypassing licenses. 'Code copying' premise collapses—moats shift to brand, community, data.
Doubao's 345M Users Start Paying — China's AI Free Era Ending
Doubao launches paid tiers at ¥68/month; users rage to trending. 345M MAU's inference costs force ByteDance to charge, exposing AI's "apologize not fi
agui Exposes AI Chat Flaw: Streaming Fails, Tool Calling Needs Unified UI Protocol
agui unifies text, tool calls, and errors into one stream. It fixes UX collapse during AI tool use, evolving frontends from typewriters to true protoc
Updating 1% Params: Fine-Tuning & Quantization Slash Custom LLM Deployment Barriers
Fine-tuning turns LLMs into specialists; quantization trims them down. LoRA updates just 1% of params, enabling SMEs to customize AI with consumer GPU
Million-Param GPT on Journey to the West: Demystifying LLMs Is the New Imperative
Training a million-param mini Chinese GPT on Journey to the West locally reflects the industry's urgent need to demystify the LLM black box and master
RAG's Five Stages: LLMs Embrace Open-Book Exams as Enterprise Standard
RAG is the enterprise LLM standard, enabling "open-book exams" via knowledge retrieval. But accuracy, engineering complexity, and data cleaning remain
Hermes Agent Automates WeChat Official Accounts—AI Drops to One Command
Hermes by Nous Research: a one-command Agent automating WeChat/Feishu bots, marking AI's shift from dev-only to office use.
LLMs Keep Hallucinating: RAG Becomes the Enterprise Standard Config
RAG makes AI check external data before answering, fixing knowledge lag and hallucinations. It's core infrastructure, not a patch, for safe private da
Cursor Opens Agent Engine SDK: AI Code Editors Grab Infrastructure Turf
Cursor's TypeScript SDK opens its Agent engine to developers with local, cloud, and self-hosted runtimes. It signals Cursor's shift from tool to infra
OpenCode Skills: AI Assistants Split 'Knowing' From 'Doing' On Demand
OpenCode Skills inject AI knowledge via Markdown on demand, splitting "manual" from "toolbox." It solves context crowding but adds maintenance overhea
RAG's Accuracy Flaw: Why Vector Databases Alone Fail Enterprise Knowledge Bases
RAG often misses the mark. Naive similarity search yields duplicates and ignores constraints. Retrieval strategy is the real watershed for viable know
Robot Vision Standard Interface Emerges: Linux Ends Custom Camera Drivers
ROS enforces camera coordinates, Linux V4L2 abstracts hardware—robot vision is standardizing. Tech barriers drop, but lock-in risks accumulate.
SKILL.md Sets AI Rules — Agent Race Shifts from Models to Process Constraints
SKILL.md: Markdown workflows for AI coders. Same model, different decisions with process constraints. Marks Agents' shift from chat tools to process e
opencode Adopts MCP Protocol: AI's "USB Port" for External Tools Starts to Unify
opencode updates MCP config guide for local and remote servers. The protocol cuts AI-tool integration costs, but data security risks and big-tech ecos
Superpowers Gives AI Assistants Engineering Brains, Ending Reckless Coding
Superpowers is an open-source AI coding skill framework packaging TDD and code reviews into skills. AI coding is shifting from "writing code" to "doin
HKU Open-Sources DeepTutor: AI Tutoring Deployment Barrier Drops Again
HKU open-sources DeepTutor with guided install and local RAG. AI tutoring deployment barriers drop, but hardware demands still block non-technical use
LangChain Dismantles Omnipotent AI: Multi-Agent Becomes Pragmatic Enterprise Choice
LangChain replaces omnipotent AI with specialized multi-agent collaboration. This cures tool-selection errors, shifting AI from tech demos to true bus
OpenHands Hits 40K Stars — Open Source Catches Up to Closed-Source AI Coders
OpenHands is an open-source AI coding agent in Docker sandbox with 40K+ GitHub stars. Open source rapidly closing gap with closed-source coding Agents
Wang Zirong's 48-Day AI App: Dev Metrics Shift From Code to System Design
Wang Zirong's 48-day AI app sparks debate. Code lines no longer measure developers. The real AI-era bottleneck is system design and engineering comple
900K-Token RAG Test: Simplest Line Split Wins; Enterprise KBs Stop Overpaying
Most enterprise RAG projects fail at chunking. Latest 900K-token benchmark: simplest line splitting is most accurate. Chunking strategy > model choice
Copilot's Token Billing Shift: AI Giants Pass the Tab to Developers
Copilot shifts to Token billing while OpenAI bleeds cash. AI tools move to usage-based pricing, making costs soar. We must reassess the ROI of AI adop
DeepSeek V4 at 1/22 GPT-5.5 Price: LLM War Shifts to Efficiency
DeepSeek V4-Pro: $3.48/M tokens (1/22 of GPT-5.5). Architecture, not subsidies, rewrites AI economics, making 24/7 Agents individually affordable.
WordPress AI Plugin Hits 12 Pitfalls: AI App Bottleneck Is Engineering, Not Models
A developer hit 12 pitfalls building a WordPress AI image plugin across 3 languages. The real AI app bottleneck is engineering toolchain, not models.
Xiaomi MiMo Offers 100 Trillion Free Tokens: LLMs Burn Cash for Developers
Xiaomi MiMo offers 100 trillion free Tokens via Claude Code. Not a tech breakthrough, but an LLM acquisition war using free compute to capture develop
Codex Directs DeepSeek Grunt Work: AI Multi-Agent Collaboration Counts Costs
Developers use Codex for decisions and DeepSeek for execution to slash token costs. AI apps shift from single-model brute force to multi-agent cost co
Claude Code Exposes 4 Sub-Agent Isolation Tiers: Anthropic Teaches AI Teamwork
Claude Code reveals 4 sub-agent isolation tiers, from coroutines to tmux. It's the core engineering challenge for AI coding shifting from solo to team
Re-embrace Small Steps: Incremental Verification Beats One-Shot AI Coding
A Node.js article breaks dev into 5 stages, proving "contract-first, small steps" discipline is safer than letting AI generate entire projects at once
90% of Enterprise AI Knowledge Base Failures Lie in Retrieval, Not LLMs
When enterprise AI fails, most blame the LLM. The real bottleneck is retrieval. Vector similarity ≠ business relevance; optimizing retrieval is the cu
7 Years of Transformer Dominance: LLM Architecture Awaits the Next Reshuffle
Transformer underpins LLMs via self-attention, fixing old algorithms' parallel and long-context flaws. Grasping it reveals LLM capability limits and b
80M Tokens for 4 RMB: DeepSeek Disk Cache Rewrites LLM Inference Costs
DeepSeek's novel architecture enables disk-level caching, slashing API costs 10x. This signals LLM inference shifting from raw compute to engineering
Anthropic Adds 83 Commands to Claude Code: Terminal AI as the New Dev OS
Claude Code's 83 slash commands turn the AI assistant from a chatbox into a terminal OS, signaling AI tools' penetration into low-level workflows.
LangChain Breaks AI Into 4 Components: Orchestration Layer, Not Just Framework
LangChain splits AI into Chain, Agent, Memory, Tool. It's an orchestration layer shifting LLMs from "talking" to "doing"—crucial for anyone tracking A
Tencent IMA: Knowledge bases that self-digest are the real moat
Tencent IMA + WorkBuddy auto-digests knowledge:提炼, links, and writes back. Organized knowledge that improves with use is the new personal moat.
Open-Source Diary Agent Echoes: AI Pivots from Doing Tasks to Managing Memory
Open-source Agent Echoes uses targeted questions to complete memories and generate reports. AI's personal role shifts from "writing for you" to "remem
75% Match: Live Platforms Prove LLMs Must Master Structured Data First
Huajiao Live's AI streamer profiling matches human judgment 75%. LLM deployment requires outputting system-consumable structured data, not just genera
Curing AI Coding Amnesia: Context Engineering Replaces Prompting as Production Key
AI coding assistants forget each new chat. Context engineering—systematically assembling AI info—now determines output quality more than prompting.
Viral AI Pronunciation Guide Exposes Chinese Tech Community's Info Gap
Viral AI pronunciation guide: Codex isn't "Code-X," Claude isn't "Cloud." Misreads expose the Chinese tech community's info gap, as AI neologisms outp
Cursor Token Guide: 80% of Bills Wasted on Context; AI Coding's Crude Era Ends
Cursor token guide: 80% of bills burned on invalid context, not thinking. AI coding shifts from crude usage to precise accounting; context management